Qwen has recently launched the Qwen3-VL-Embedding and Qwen3-VL-Reranker models, which represent a significant advancement in multimodal information retrieval. Built on the robust Qwen3-VL foundation models, these new tools are specifically designed to handle a variety of inputs, including text, images, screenshots, and videos. They excel in tasks such as image-text retrieval, visual question answering (VQA), and multimodal content clustering, offering developers a comprehensive solution for cross-modal understanding and retrieval. One of the standout features of these models is their unified representation learning capability. The Qwen3-VL-Embedding model generates semantically rich vectors that encapsulate both textual and visual information in a shared space, enhancing the efficiency of similarity computations and retrieval across different modalities. Complementing this is the Qwen3-VL-Reranker, which refines the retrieval process by taking a (query, document) pair as input and producing a precise relevance score. Together, these models implement a two-stage approach that significantly boosts retrieval accuracy while maintaining exceptional multilingual support across over 30 languages. The architecture of the Qwen3-VL-Embedding employs a dual-tower design for independent encoding of inputs, while the reranker utilizes a single-tower architecture that facilitates deep inter-modal interaction through cross-attention mechanisms. This design not only enhances the models' performance in large-scale multimodal retrieval tasks but also allows for flexible integration into existing developer pipelines. The models support quantization and customizable instructions, making them practical for real-world applications where cross-lingual and cross-modal understanding is crucial. Evaluation results reveal that the Qwen3-VL-Embedding model achieves state-of-the-art performance on multiple benchmarks, particularly excelling in image, visual document, and video retrieval tasks. Despite showing a performance gap in text-only retrieval compared to its predecessor, the Qwen3-Embedding model, the overall capabilities of the new models mark a significant leap forward in the field of multimodal retrieval systems.